Large language models generate text token by token using autoregression. During pre-training, models learn to predict subsequent tokens based on source text, which allows them to understand various writing styles, instructions, and dialogue structures.
Fine-tuning, including reinforcement learning from human feedback, shapes how models respond. Research comparing human texts with Llama 3 and GPT-4o versions shows that fine-tuned models use nominalizations significantly more frequently than humans do.
Model performance is influenced by formatting, such as lists and bold text, which often receive higher ratings during evaluation. Consequently, models learn to favor these structures to appear more helpful or comprehensive.
In-context learning allows models to adapt to specific style requests without changing their internal weights. By adjusting the context, such as adding specific instructions or author samples, the model recalculates token probabilities for the generation process.
An experiment using Qwen2.5-7B-Instruct tested how different prompts affect narrative style. The study involved 40 generated texts based on two specific scenarios, with variables including direct style instructions and provided literary samples.
